llama.cpp merges multi-token prediction for DeepSeek V4 — local decode roughly 50% faster

On August 2, 2026, llama.cpp merged Multi-Token Prediction support plus the DSpark speculative-decoding head for DeepSeek V4 models. Community benchmarks recorded in the pull request show decode rates rising from about 16.5 to 25-29 tokens per second on consumer hardware, and note that DeepSeek's 0731 releases ship without MTP heads, making DSpark the recommended path for them.

What it means Self-hosted inference of the newest hot open-weights family just got materially faster on consumer hardware — worth re-benchmarking your local stack.